🍥｜anime

[nova remnant]

a pretty girl with brown hair wearing casual outfit mowing the lawn.

if it wasn't for context i woulnd't settle w/ the image quality lol

[whole portal]

oh yes, i think that vae-ft-mse-840000 works better for LCM than some other anime specific loras

but you can experiment

let me know if you find one that works better

[nova remnant]

i can't use that vae

im using sdxl

sdxl has one VAE

[whole portal]

i am talking about 1.5

has two

🙂

i am using the other one

[nova remnant]

but im thinking of going back to sd1.5

i miss a lot of the images with sd1.5

i had decent models to play with too

only if there was a way to apply the sdxl coherence on sd1.5

[whole portal]

https://civitai.com/models/152040/xlvaec

i am using this one for sdxl

it gave me pure red where 0.9 could not 🙂

[nova remnant]

hmm

but... NansException
ENGLISH
experimental VAE ・Encoder is not adjusted ・Refiner is not used for images ・NansException

[whole portal]

i know, but it usually works fine

it is "use in case of emergency" 🙂

when everything else fails 🙂

[nova remnant]

ahh ok

but i dont have issues with sdxl vae

[whole portal]

found it

#🍥｜anime message

first one is wirt 0.9

second with this one

drastic difference in red

stronger lcm will fix also fix the red, but i was not aware that strengt is too low at that time

[nova remnant]

used it with sdxl render

also on sd1.5 its nice to render at 768x1024 w/o deformity

[terse ore]

Any decent upscale workflow for comfyUI? My stuff looks terribad

[wispy canopy]

inching closer and closer to a real final merge

now, gotta go back and see if i didn't fuck up perspectives on landscape images

[nova remnant]

its mental lunacy how much deformity occurs in sd1.5 on a regular basis

[wispy canopy]

wow, having the hands from bpn1.2 back feels soooooo good

so much less worrying if the hands are borked

[nova remnant]

expressions are limited from how much you have to hide obvious deformity

cant make her hold apple just like that

[wispy canopy]

hands. fixed.

[nova remnant]

easy way to add expression to character is to use dynamic expressions

but adding dynamic pose can lead to weird posture

[wispy canopy]

naaah ^^ works like a charm

you'd better specify WHICH pose you want

i personally like 'confident stride'

[nova remnant]

well i meant the actual word dynamic pose

its like adding an easy cinematic like effect

try with weight

im not using the word anime in the prompt

which leaves me a lil confused cause this model can also generate realistic images

[wispy canopy]

mine IS a generic model so if i wanna make anime i'll have to add it 😛

[nova remnant]

a racoon driving a car along side a beach resort

[wispy canopy]

both mecha waifu at the same time 😄

[nova remnant]

lol

[nova remnant]

portrait of a girl with intricate hair ornament and assorted colorful gems

a bald old man in rugged clothing wielding a spear

didnt wield too well

a woman sits at a cafe, indoors, with her laptop and phone

anyone wanna try that on sd1.5 ?

[bitter latch]

a woman sits at a café, indoors, with her laptop and phone, sharp focus, elegant detail, HDR, detailed shadows, cinamatic lighting, unreal render, realistic

[nova remnant]

neat but did you have to use anything extra to get the hand right?

[bitter latch]

used negatives , sampler restart

[nova remnant]

ahh ok, i was trying out sd1.5 some mins ago and i was getting so many weird forms

the one i was trying out is about a girl holding an apple

[bitter latch]

lora:LCM_LoRA_Weights_SD15:1, midjourney style, a woman sits at a café, indoors, with her laptop and phone, hypernet:mjv4Hypernetwork_v1:0.8octane render
Steps: 20, Sampler: LCM, CFG scale: 4, Seed: 2945522132, Size: 768x768, Model: ElementAOM3Hello_merge, VAE: clearvae_v23.safetensors, RNG: NV, Lora hashes: "LCM_LoRA_Weights_SD15: aaebf6360f7d", Version: v1.6.0-2-g4afaaf8a

that came out.. umm interesting

[nova remnant]

civit ai is screwing up their web pages with javascripts for xmas decoration

cant use some of the menus

[nova remnant]

lora weight might be too high

[bitter latch]

ah ha

[nova remnant]

try this ....

A brave girl soaring through the skies on a majestic dragon, exploring uncharted lands and befriending mythical creatures.

[bitter latch]

turned it down to .75 and just did midjourney style , with my soft anime style I have saved. replaced laptop and phone with cat 😄

[nova remnant]

nice.... 75 is ok i think, i usually keep the weight to 0.6.... while using lcm sampler

[nova remnant]

dark contast

but it got the context

[bitter latch]

well that went wacky

[nova remnant]

heh

[bitter latch]

tho I bet it would do better with my fuzzyanime merge

[nova remnant]

A girl with a paintbrush in hand, creating a vibrant masterpiece on a blank canvas, surrounded by a messy art studio.

A girl with a mischievous grin, riding a bicycle through a bustling cityscape.

[bitter latch]

A girl riding on a Pixar Dragon over and epic fantasy land

haha

ahh I see

[nova remnant]

A young witch with flowing pink hair and shimmering emerald eyes brews a mysterious potion in her dimly lit, cobweb-draped apothecary. The air is thick with the scent of herbs and spices, and bubbling flasks glow with an eerie luminescence.

[nova remnant]

a beautiful young woman with flowing crimson locks and piercing scarlet eyes. A sly grin graces her lips

google bard is struggligng to keep up with my prompt demands

[bitter latch]

the what?

[nova remnant]

glitching out in the middle of displaying them, like 3/10 and bonk

[bitter latch]

ahh

[nova remnant]

A gifted musician girl plays a beautiful melody on her violin, her notes filling the air with magic.

A creative and imaginative girl spends her days drawing and painting, her world filled with vibrant colors and fantastical creatures.

[nova remnant]

an old man with a large bag on his back climbing uphill as the sun sets in the horizon.

[bitter latch]

personally I like playing with the settings and coming up with my own prompts and methodology of settings and such

or starting with someone elses prompt and redoing it 😄

[bitter latch]

an old man Anthro Fox with a large backpack on his back climbing a mountain as the sun rises over the horizon. (pixar style)

[bitter latch]

someone forgot pants haha

[nova remnant]

yeah .... :p puss in boots in an enchanted forest teeming with flowers and plants

[bitter latch]

eular A with default no other styls

[bitter latch]

I love this furry and anime merge I made

does real good

twins?

[bitter latch]

oooh

[nova remnant]

@eface you may not have to use low quality, worst quality wtih https://civitai.com/models/119032/unaestheticxl-or-negative-ti

[steady grail]

it feels so natural

[nova remnant]

hi .. yeah sdxl is so much more fun

[terse ore]

But doesn't sdxl have more limited style compared to all those 1.5 checkpoints?

[native halo]

yes,its up to you to decide which one to use

[nova remnant]

styles? no, with SDXL you can have much more dynamic styles, but if you mean community models then yes, because SDXL initially started as a huge resource hog which made it difficult to train and use, but that's changing now.

[terse ore]

I mean like, 1.5 has realism, flat 2d, 3d, 2.5d, monochrome, etc

[native halo]

u need loras for that

and its expensive to train on XL

[nova remnant]

sure but key is coherence

[terse ore]

Oh, Lora for sdxl. I see that makes sense. I never got into sdxl but that makes sense

[native halo]

yea all the 1k+ artist loras for 1.5 are not on XL because money and difficulty to train

[terse ore]

I was trying to make a coherent realistic crowd but gave up fast on sdxl. So I will have to try sdxl for that soon

[native halo]

NovelAI probably has all those styles for their XL private model because they are big company with access to clusters of A100's

[nova remnant]

well you can use SDXL better than before because of it becoming less of a resouce hog but despite that its still a bit high ended to train which is why you dont see that many contrintuion from the community like you see with sd1.5, but sdxl has much better coherence level that sd1.5 doesn't so that's a foundational differnce bewteen the two.

[terse ore]

I see. I suppose it depends on what one uses it for then. If it's available for sdxl then it would be best to use that, all else being equal

[nova remnant]

they are not equal per se

sd1.5 was more accessible and sdxl is more dynamic

[native halo]

yea if u want certain styles or to prompt certain anime characters u go for 1.5 if u just want some generic anime img with good coherence u can try XL

[nova remnant]

@native halo what if I want both and bodysuits

[nova remnant]

you could train sdxl to focus on a particular anime theme too, its that the community hasn't delved into beause of training resources needed

Hello chatorqx

[native halo]

(latex bodysuit,skin tight:1.5)

[native halo]

yes u can novelai did it with their XL release

but no more leaks from them

[nova remnant]

oh im sure things will eventually come in reach for average consumers when community starts to build on sdxl

[native halo]

we can do it right now just need money

💸

[nova remnant]

lol money has been one of the biggest hindrance or slow growth for technology and society's progress

if a good idea doens't cash in it gets rejected

also generosity in this economy is counter to survival

[nova remnant]

🙂

[nova remnant]

i wonder when SD will reach a level where users can generate gif files on the fly

gif animations

[native halo]

maybe next year cause animatediff still eats a lot of vram

[nova remnant]

YES

prolly

i hope

[nova remnant]

big tooth monster

[native halo]

that chin

nai has really great hands on their XL model

[nova remnant]

where are you generating them?

[native halo]

its from novelai twitter

[nova remnant]

oh

[native halo]

i thought they used A100 cluster but its actually H100's cluster 🤠

[nova remnant]

isn't h100 even more powerful?

[native halo]

yea it was like 30k$ each

[gusty palm]

playing around with CADS some more i love the contrast

[nova remnant]

whats a CADS?

[gusty palm]

https://github.com/v0xie/sd-webui-cads.git

i cant pretend to understand it

[nova remnant]

hmmm so that adds more effects?

its basically adding noise to change image outputs to look more cinematic maybe

[gusty palm]

it does somethign to the noise i think

[nova remnant]

CADS greatly increases diversity of generated images by adding scheduled noise to the conditioning at inference time.

something along the line of using detail lora

maybe it's method is better idk just guessing

[gusty palm]

i think it works in a similar manner to cd tuner

but also not like it at all

[native halo]

yea cd tuner adds a weird noise to the img that i dont like

like static

[gusty palm]

i had that problem initially but i didnt have it last time i tried it

[native halo]

it only happens when u increase contrast to the max

[nova remnant]

have you looked at how SDXL generate realistic faces?

i mean in sd1.5 there is a pattern to face based on partiuclar training method which can be common for different models

but sdxl does it so good you won't be able to tell if she is real or artificial

unfamiliar but familiarly gorgeous

[gusty palm]

i havent really tried xl at all yet tbh

[native halo]

never had that problem

[nova remnant]

that's a terribly cute face

[native halo]

oh u mean realistic? depends on the model

[nova remnant]

its bit different with sdxl

it doesn't have a fixed common pattern

fibonacci ... unending variation with unique symmetry of perfection

[steady grail]

this is a little risque, or maybe it's just me

maybe i'm wrong

[nova remnant]

spoilered it for that reason, but she is wrapped up in towel

[steady grail]

lol

[nova remnant]

beauty can be risque

and possibly because how much society is held away from it

[native halo]

thats why we need to eat vegetables so we can keep a healthy body

[steady grail]

let's hope the staff is of same advise lol

[nova remnant]

how on earth is your character name blending with the background?

[young fjord]

invisible text

᲼᲼

[nova remnant]

are you like a super shadow admin?

[young fjord]

no

[nova remnant]

how did you do it?

[young fjord]

copy that

and paste

[nova remnant]

᲼᲼ cant see anything

[young fjord]

᲼᲼᲼᲼᲼᲼᲼᲼lol

[nova remnant]

can't see

[young fjord]

᲼᲼᲼᲼it's fun to play ᲼᲼᲼᲼with

[nova remnant]

are you discord dev?

[young fjord]

...no

[nova remnant]

im watching a chiropractic asmr

this chiropractor cracks pretty girls to relieve pains

[nova remnant]

@native halo big warrior to protect us

[native halo]

i like that big massive sword

[sharp swan]

https://tenor.com/view/the-rock-think-the-rock-meme-gif-24810091

[primal wharf]

https://tenor.com/view/cheekibreeki-cheeki-breeki-intensifies-stalker-bandit-radio-stalker-clear-sky-gif-24295015

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1180759244371402792/00191-2468585765.png?ex=657e96bd&is=656c21bd&hm=f6a18cdeb919ac92be28d65cf413c9a6d2e9dc8181152c941605cbc68709975c&=&format=webp&quality=lossless&width=441&height=662

[gusty palm]

ik this gets asked alot but is abstract geometry art:1.2 the same as (abstract geometry art:1.2)

i guess i could test

[gusty palm]

link

also u should check out myles power video about chiropractic

[gusty palm]

can't remember if I posted this yesterday

[gusty palm]

thats cool looking

if power rangers action sequences were in that style id watch it

[granite shell]

is a detailer broken lol its messing up all my gens

[gusty palm]

giant waifu 😳

how do i make it make my characters smoking we'd

would i need a lora for that

[supple raptor]

nope, without the () around it, the :1.2 is just treated as text and has no influence on the emphasis.

[gusty palm]

ooh

CADs upscale vs no CADs upscale using img2img
cant decide which i like more

0.36 denoise

[gusty palm]

how about the difference between abstract and (abstract) guessing paranthesis makes it stronger but nott sure if theres no numbers

[vital raptor]

not having the numbers is the same as having 1.1

(abstract) is the same as (abstract:1.1)

at least in automatic1111, if you use a different interface it could be different

[supple raptor]

To expand on this each set of parenthesis multiplies the emphasis by 1.1 if no numbers are used. So:
(abstract) = (abstract:1.1)
((abstract)) = (abstract:1.21)

[bitter latch]

with lms

[toxic dragon]

Some quick style testing.

[bitter latch]

LCM loves this style too

hmm missing an ear?

hmm looks like a anime 80's vibe record poster

[vital raptor]

reminds me of something you'd see on a synthwave youtube thumbnail 😄

[wispy canopy]

show me ur best waifu!

[bitter latch]

well "synth anime wave vibes" was part of prompt

[toxic dragon]

More of this style because I love it.

[nova remnant]

sends it in private

[bitter latch]

surprised LCM is doing so well

without the lora

it's 6am I should be in bed!

..sigh

with lora

[nova remnant]

i will message you the link, its safe but since not relevant to anime

[nova remnant]

ok i will check but im only watching chiropractor video that i messaged you cause of cute girls, not chiropractor itself

[nova remnant]

t2i adapter for sdxl doesn't have openpose

nvm i see it

basically all i need - canny, depth and openpose .. 150mb each for sdxl 🙂

[vital raptor]

Anyone have any XL waifu models to recommend? I need to find some good ones to test my lora with

[native halo]

need to be more specific,like whats your lora or what style u lookin for

[vital raptor]

my lora is basically a character/face I trained for realistic models, but I was just doing a x/y with all my models and realized I only have a couple anime related models. Something clean 'normal' anime would be nice, but I'm not being picky

This is my lora on copax

[native halo]

for XL id say Himawarimix or Blue pencil those are very good

[nova remnant]

a waifu xl model https://huggingface.co/hakurei/waifu-diffusion-xl

yeah and bluepencil is good too

[vital raptor]

This one is counterfeit

same prompt and everything

[nova remnant]

countefeit is popular too

[native halo]

nekoXL is also good

[nova remnant]

i use bluepencil mostly

checking out nekoxl

[native halo]

only way to know which one u prefer is to download them all and test them

[vital raptor]

yeah that's usually the way i do it, just figured I get some opinions for starters 😛

[nova remnant]

lol i dunno if i got the right one for neko xl

[native halo]

there are 2 nekoray xl and nekoxl the 2nd one is nsfw

[nova remnant]

you know what... i just saw you are absolutely correct

another one you could try is reproduction

[vital raptor]

yeah, I like blue pencil too, it added a third person for some reason though 😄

[vital raptor]

the lora i made I've mostly been testing on a merge i made with blue pencil and betterthanwords

[nova remnant]

another Animagine XL

this should be good too https://civitai.com/models/117259/anime-art-diffusion-xl?modelVersionId=128592

i havent tried it out but the author has great models

also im waiting for them to come out of alpha and beta

[vital raptor]

hmm, i had alpha2 of that and wasn't getting great results, i'll have to try the newer one

[nova remnant]

sounds familiar

but the author has great work given we are not using alphas

those are basically incomplete

what i mean by great work.. im refering to sd1.5

dreamshaper has been a sensation.. but their xl version is stil in alpha too

i loved using it on sd1.5

[vital raptor]

yeah, it was pretty good on 1.5

[nova remnant]

yep very versatile

[vital raptor]

looks like the rest of these anime models need some different prompting to get reasonable results, or my lora doesn't work with them, the ones i posted were using the same prompt from my realistic model testing

[nova remnant]

they all lack training to a degree

some are better than the other but they all could use better training which hasn't been possible due to sdxl resource demand

[vital raptor]

not surprising, sdxl is a pain to train

[nova remnant]

but so far bluepencil is fairly good, given they merged like 100s of models

you can see for yourself 🙂 https://huggingface.co/bluepen5805/blue_pencil-XL

List of models (not all are included)
merge history - v1.0.0

[vital raptor]

that's a stupid amount of work

[nova remnant]

yes lol

gotta appreciate that

[vital raptor]

yeah, animeArtDiffusion alpha3 is way better than alpha2 was

😮 nekoray needed CFG15 to just be not washed out, i like it though

[nova remnant]

have you tried the new LCM lora?

[vital raptor]

haven't heard of it

[nova remnant]

https://huggingface.co/latent-consistency/lcm-lora-sdxl

basically it reduces the need for high sampling steps and cfg and still deliver enough noise to generate decent images

and that reduces the render time greatly

its a new tech that has come up in the last 2 weeks

also making use of SDXl lot more accessible

[vital raptor]

is this the same as the SDXL Turbo thing or just a similar idea? I've heard about it but haven't looked into it

[nova remnant]

no

turbo is another innovation as of late

[vital raptor]

i see

[nova remnant]

they are different but can work together

turbo model requires you use the turbo model

but lcm lora works on all checkpoints and there are seperate lcm lora for both sd1.5 and sdxl

and point to note, turbo sdxl is 6gb .. lcm lora is few hundred mb

380mb

[vital raptor]

interesting

[nova remnant]

my typical workflow setup using LCM LoRA

comes out with that level of image quality

sunny day, puffy clouds, beautiful palm trees, sea beach, and a close up of a pretty girl <lora:lcm-lora-sdxl:0.6>

[nova remnant]

there is a LoRA for turbo sdxl too, about 700mb size which you can use with any checkpoints, and its based on turbo sdxl model

https://civitai.com/models/215485

LCM LoRA and SDXL Turbo are both LoRA (Low-Resolution Activation) models for Stable Diffusion that can significantly speed up image generation. However, they have different strengths and weaknesses, making them suitable for different use cases.

LCM LoRA is a more general-purpose accelerator that can be used with any Stable Diffusion model. It is particularly well-suited for generating high-resolution images (1024x1024 pixels or larger), where it can achieve real-time generation speeds. However, LCM LoRA is not as fast as SDXL Turbo for generating lower-resolution images (512x512 pixels).

SDXL Turbo is a more specialized accelerator that is specifically designed for the Stable Diffusion XL model. It is the fastest LoRA model available for Stable Diffusion XL, and it can generate 512x512 images in real time. However, SDXL Turbo cannot be used with other Stable Diffusion models, and it is not as well-suited for generating high-resolution images as LCM LoRA.

Here is a table summarizing the key differences between LCM LoRA and SDXL Turbo:

Feature	LCM LoRA	SDXL Turbo
Compatibility	Any Stable Diffusion model	Stable Diffusion XL only
Resolution	High-resolution (1024x1024 pixels or larger)	Lower-resolution (512x512 pixels)
Speed	Real-time for high-resolution images	Faster than real-time for lower-resolution images

Use case recommendations:

• If you need to generate high-resolution images (1024x1024 pixels or larger), then LCM LoRA is the best choice.
• If you need to generate lower-resolution images (512x512 pixels) and you are using the Stable Diffusion XL model, then SDXL Turbo is the best choice.
• If you are using a different Stable Diffusion model, then LCM LoRA is the best choice.

In general, LCM LoRA is a more versatile and powerful accelerator, while SDXL Turbo is a faster and more specialized accelerator. The best choice for you will depend on your specific needs.

[vital raptor]

just tried them both, LCM works better with my model/lora combo so far.. but that's impressive

[vital raptor]

hmm, 8.5s with a step of 2x hires fix

[nova remnant]

what was your sampling steps

and cfg?

[vital raptor]

8, 1 hires, cfg 2, denoise 0.25

used blue pencil

[nova remnant]

and basic canvas size?

[vital raptor]

1024x1024, 2x hires

[nova remnant]

you have rtx card

[vital raptor]

4090

[nova remnant]

ahh nice that figures

mine is 3060 wtth 12gig vram and 16gig ram and i have to avoid hires upscaling while using sdxl

you can try a resolution like 768x1024 too if you don't like square

[vital raptor]

normally i would limit it a bit just because the speed, but only needing 1 step is crazy

[nova remnant]

yeah

btw if your sampling steps is set to say 8 you could try 4 for hires steps

cfg should be good with 2

[vital raptor]

oh, i was mistaken, that's 2 hires steps.. any higher and it breaks the image, their faces start going sideways 😄

[nova remnant]

hmm

it shouldnt distort the image with 4 hires steps

unless there is some lora interfering

[vital raptor]

it might be my trained lora

[nova remnant]

ahh could be it

i've tested hires with 1024x1024

only reason i dont' usually use hires in sdxl cause of ram

[vital raptor]

yeah, it's fine without my lora at 4 steps, adds 2 seconds to the render though

[nova remnant]

ahh ok

[vital raptor]

2 and 4 steps without the lora

changes the image a little still but doesn't mess them up

[nova remnant]

they came out nicely

[vital raptor]

those both rendered with CFG 1 for some reason

I think i accidently scrolls the slider 😛

[nova remnant]

1-2 cfg is good

[vital raptor]

1.5 does pretty good with my lora, 2 gives it more of a nice contrast punch though

[nova remnant]

hugging face also recommends turning off cfg but you cant do that in a1111 .. but also i thinki like the cfg 2 for the image quality

[vital raptor]

yeah, 2 seems to work well for my taste

[nova remnant]

yeah i think it adds to more detail

in case you get too strong contrast try the lcm lora wieght to something like 0.6 and use LCM sampler

if you dont have LCM sampler you can easily get it by installing the extension AnimateDiff from extension tab

[vital raptor]

i did these all with the lcm sampler

[nova remnant]

ahh sweet

0.6-0.8 lcm weight is fairly good

1 is bit too strong

[vital raptor]

I've spent the whole day staring at burned images while i tried to get this dumb lora to train, i'll have to look closer tomorrow when my eyes have rested 😛

[nova remnant]

ha.. get some rest

[vital raptor]

i should...

[nova remnant]

its easy to get overly absorbed playing with SD lol..

[vital raptor]

yeah, i've been messing with it for quite awhile but just recently decided to try to move to sdxl

it was way easier to custom train my 1.5 model

[nova remnant]

i moved to sdxl 3 days now i think, i mean i tried it before lcm lora too but din't feel like using it for slow render

[vital raptor]

yeah, i tried it when 1.0 first came out, but my 3080TI was to slow 😏

[nova remnant]

what did you do with 3080 ?

[vital raptor]

it's currently sitting in my home server running an LLM i've been tinkering with

might end up selling it though, prices are crazy right now

[nova remnant]

lol i see, thats a good card too

does that have 16gb?

[vital raptor]

12gb

[nova remnant]

ok

i didnt have any good deal with 16gb so i had to settle with 12gb rtx

btw how are you managing your negative prompt with sdxl?

[vital raptor]

for the most part I haven't been using it much since i was having some problems with it

[nova remnant]

ok i wonder what problem you mean, but usually you don't need much generic prompt like you did with sd1.5

[vital raptor]

but those LCM tests had "noisy, blurry, amateurish, sloppy, unattractive," in the negative

[nova remnant]

you are using lcm lora in negative?

[vital raptor]

I was having problems with my loras burning whenever i put anything in the negative

no, i just mean when i generated the images earlier

[nova remnant]

oh

ok well lcm lora goes in positive box, you can leave negative empty but i use two recommended TI

https://civitai.com/models/119032/unaestheticxl-or-negative-ti

just these two unaestheticXLv13 unaestheticXLv31

[vital raptor]

yeah, i have those

[nova remnant]

nothing else in negative

[vital raptor]

haven't messed with them much, I'm trying to stay away from them if i can

[nova remnant]

those TI adds a subtle nice touch

you could also just use the Sky version for most anime too

31 and 13 are slightly different in their qualities

i have a ghost in discord

i hear beeping but i see no visual indication where its coming from

[vital raptor]

wait... TensorRT works with LCM..

[nova remnant]

i havent tried it with lcm

are you using it?

[vital raptor]

yeah, just finished generating the rt engine for bluepencil

4 hi res steps now takes 8.8 seconds instead of ~10.5

[nova remnant]

did it make extra 6gig tensor model?

[vital raptor]

i didnt convert lcm with it

[nova remnant]

usually whenever you make a tensor rt profile on a model that creates huge file just for that profile

[vital raptor]

yeah, for the model, i forget where it even puts those

[nova remnant]

and if you were to swap several different modesl for your work then creating tensor rt profile for each of them takes up lot of space

[vital raptor]

yeah, bluepencil is a 5gb TRT file

[nova remnant]

its cool if you have large ssd

in my case im with 500gb ssd

[vital raptor]

i have 8TB in this computer

[nova remnant]

oh damn.. thats ssd?

[vital raptor]

3 2TB NVME, a 1tb nvme and a 1tb ssd

[nova remnant]

awesome

[vital raptor]

yeah, i splurged when i built this machine a few months ago

[nova remnant]

mine is nvme too, its a small card that goes in mobo

[vital raptor]

got a motherboard that could fit 4 nvme and filled them up

[nova remnant]

lol nice

my next upgrade im looking at is to get more ram and larger storage

[vital raptor]

I should go through and delete some stuff from my AI folder though

[nova remnant]

isnt that 800gb tho?

[vital raptor]

lol yeah

[nova remnant]

ahh i thought yousaid 8tb

[vital raptor]

i did, this is what my AI folder is using

[nova remnant]

oh

[vital raptor]

490GB of that is models/loras and training datasets

[nova remnant]

🙂

curious about your inovlvement if you are participating in releasing sd content publicly

i do all this for fun as a user

[vital raptor]

not really, just messing around as a hobby

[nova remnant]

ahh ok me too

[vital raptor]

if I ever ended up making something more general use i might release it, but my models are trained with a specific face, not super useful to anyone else

besides that I post random images on twitter sometimes 🤷‍♂️

[nova remnant]

nice

speaking of posts ... its good to share online, i upload selective ones on my civitai profile too

[vital raptor]

yeah, i actually say some of your posts after you posted them here when i was looking at models 😄

[nova remnant]

ahh ok, i do that to share metadata lol

[vital raptor]

yeah, I haven't shared anything on there yet since i'm using a private model my data isn't as useful

i guess prompting ideas are still useful

[nova remnant]

but it could be something useful for people to see your workflow when you share metadata

i used to dump all the rendered images on tumblr, i still do, but i like how civitai adds all the detail of an image

[vital raptor]

i'll probably start posting stuff on civitai that I don't put on my twitter

[nova remnant]

should be good 🙂 cause twitter doesn't share metadata

[vital raptor]

yeah, and my twitter is family friendly because of friends kids and stuff, so there's a lot I wouldn't post there

[nova remnant]

i might asl well clean up my ssd and upload all the images to free up space

yeah i get what you are saying

with civitai i also find it convenient to look up prompt ideas that i came up with

especially when im not saving the image in my drive

[vital raptor]

yeah, i always have it open incase i need something

[gusty palm]

how do i convery my current model to xl

[vital raptor]

you can't, they have to be trained from scratch

[gusty palm]

booooooo

[vital raptor]

i agree

[gusty palm]

i bet sdxl cant do this

[vital raptor]

if you find the right models and/or loras to merge you might be able to manage it, or with enough images generated from your existing model you could train a SDXL lora/model to do it

[gusty palm]

oh hmm

its ok i cant really run xl models afaik anyway and the look is mostly due to lora

[vital raptor]

i will say that training sdxl is a bit of a pain though

i really like that style though, they're adorable

[gusty palm]

ty

i want like, a 9 by 6 blacklight stoner poster of em

[nova remnant]

drugs bad

get help

[nova remnant]

inspired by some cat bot gens by chance ? :p

[nova remnant]

feeding keywords to bard 🙂

for the sake of idea.... the process involves coming up with random keywords you can think up, then ask bard to write a creative prompts based on thsoe keywords, then give it a final touch to the prompts by editing the phrasing

[gusty palm]

i wqas doing something similar to that with the oobagooba text gen

[nova remnant]

nice 🙂

[gusty palm]

i bet bard is better though i was never very happy with the results from the one i tried

bard always said my prompt was too lewd

or well c.ai did

[nova remnant]

bard is less strict on censoring lol but it wont interact on nsfw content

but if you say the word binkin to bing or chatgpt .. you get a 5 line lecture right before it shuts down

[gusty palm]

lol

this is a christian chatbot please no exposed ankles

[nova remnant]

ankle arches are one of the most beautiful aspect of female anatomy

why ankles

cause grace

the supple arch

[nova remnant]

i sometimes have to wonder in awe how nature's beauty flows through girls

[gusty palm]

amazing what god does

[nova remnant]

yes althoug im yet touch god lol i try to look for tangibility

[gusty palm]

as long as u dont actually look or enjoy it, thats not allowed

[nova remnant]

but im sure there is a deep natural phenomenon behind creation

[gusty palm]

i needa learn how to fix minor details

[gentle musk]

3 steps

[nova remnant]

think of 5th dimension from the perspective of our 3 dimension it's not cogntively comphrensive

[gusty palm]

yeah even 4th dimension is pretty much impossible

the analogy of a 2d creature not understanding what a 3d world looks like helped me understand that

[nova remnant]

yes time that shoots on a line we can't even see it at a glance from past to future

[nova remnant]

yes, the 2D flat world 🙂

carl sagans famous explanation

[gusty palm]

like a 2d creature with only ledngh and width wouldnt be able to understand height it would just look like a morphing line segment to them

theres this like, pretty blond lady on youtube who does little short explanations too

[nova remnant]

yes exactly

[gusty palm]

using paper cutouts and a whiteboard

[nova remnant]

just as much we are incapable of understanding what's beyond our 4d world

can i have the link to that video please?

[gusty palm]

its crazy that we can simulate 4+d though

just by adding another vector or whatever

sure

[nova remnant]

yes, and also theorize what's farther beyond

[gusty palm]

https://www.youtube.com/shorts/VwofJ3wkzn8

[nova remnant]

watching

[gusty palm]

geometry asmr

[nova remnant]

that's a beautiful analogy

[gusty palm]

right>?!

and alil creepy id like a sci fi movie about those

idk if thats the plot of stranger things or not

[nova remnant]

i think one of the biggest cause for our denial is that the danger of assuming things that could be detrimental

[gusty palm]

and no evidence

[nova remnant]

i think its very important we find tangibility in seeking mysteries of creation

[gusty palm]

maybe if we were finding empty unbroken eggs

[nova remnant]

inside i know there is god like creator but not by definiton of religion tho, cause how they exploit but there is definitely deeper meaning to reality

and i can't convery my subjective view for objective understanding, its not possible, and that's one of the areas its hard to study consciounses

[gusty palm]

i wont say to know for sure, ive been watching a lot of debates tom jump aka tjump has a good way of explaing epistomology

[nova remnant]

interesting

any video link on that?

[gusty palm]

it'll be tough for me to think of a quintessential tjump video but I have a clip somewhere where he explains his logic and it seems really solid to me I'll try to find it or a good intro video for his channel

[nova remnant]

ok

you have me curious

[gusty palm]

he explained determinism in a way that blew my mind pretty well

[nova remnant]

im yet to understand the whole deal with determinism

i have mixed view of it from what's out there

[gusty palm]

baw she's cold

she should take her fire inside

[latent vector]

undetected ai art :P

time to steal a job

[nova remnant]

Don't forget to click Correct

idk how those eyes passed through lmao

[latent vector]

we are rising

[primal wharf]

Cheeki breeki iv damke

[wispy canopy]

FUCK

was making new model merge merge, decided to check vae... get different output on known good vae

:/

[vital raptor]

I hate that

[wispy canopy]

it's extremely minor tho

[vital raptor]

SDXL, 1024x1024, 8 steps, CFG Scale 1, hires fix x2 (2048x2048), 2 steps, using LCM and TensorRT. 7.3 seconds render time 🥰

[wispy canopy]

is probably going to be ok

now it's just a matter of making more preview images

[gloomy scroll]

1girl white oversized hoodie

[wispy canopy]

https://civitai.com/models/148871 WOOOOOO new model finally out

thx @untold glacier for editing the title image ❤️

[untold glacier]

🥳

[nova remnant]

It's you

Back from the dead

[untold glacier]

https://tenor.com/view/shush-shh-anime-your-lie-in-april-silence-gif-15792775

[latent vector]

pixel art

[onyx girder]

Sapo banguela

[tropic wraith]

@onyx girder

FAQ: I'm new here! How do I get started?

Welcome! Start by heading over to #1072220168534642768 to get yourself situated and help find the channels you are looking for! Please make sure you review our #✍🏼｜rules-and-tos and feel free to assign yourself some #👥｜roles as well! Answer any questions your may have at our #1072229020520947753. There are many ways of accessing Stable Diffusion, take a look at #1080946152318443610 to start your journey!

[untold glacier]

Love this one

[untold glacier]

P R A I S E T H E H A N D

[nova remnant]

for SDXL embedding reference

https://civitai.com/models/119032/unaestheticxl-or-negative-ti
https://civitai.com/models/148131?modelVersionId=165259

[steady grail]

https://tenor.com/view/praise-the-sun-dark-souls-gif-5640777504984441459

[nova remnant]

whats going on? is there a big SD update coming?

[native halo]

yes they will prune XL again to remove more things from it

[untold glacier]

What

[native halo]

call it sdxl turbo ultra light mega 1.5 mode

[nova remnant]

really?

oh damn .. thats exciting

when is it due?

[native halo]

yea they want to create a model thats similar to 1.5 but science cant go that far yet its impossible to recreate 1.5 right now

[nova remnant]

so they are trying to optimize SDXL instead to let more people access it?

[untold glacier]

Ignore the resident trolls

[native halo]

yea just listen to the mod here

[nova remnant]

you two are confusing me now lol ..

[native halo]

just playin around

[nova remnant]

you got me all hyped up for it

but what about hands @untold glacier

[native halo]

https://media.discordapp.net/attachments/855935982943993876/1176768301880311818/ezgif-5-8d87876fc2.gif?ex=657011e2&is=655d9ce2&hm=6d22a7346f72f502ac0f55f25a136a334fabaa3a31e839e8907dc859b1d2ad83&

[untold glacier]

It was a good gen.

[nova remnant]

lol i see

and i thought something was up with SD release

[native halo]

maybe SDXL TURBO LCM ultra light ait tensorRT sdfast version 2.0 coming up

[untold glacier]

Using BPN 1.3, Eface's new model upgrade. Fantastically creative, follows the prompt like nothing else, and is quite cohesive.

SDXL of course.

[nova remnant]

interesting, i will try to download and check it out

after i get some sleep

💤

its downloading..

at present i got like 2xl model that i use

let me remove that wall of text and paste that in notepad

SDXL upcoming changes

A1111 changes ...

[keen hollow]

evenin'

[nova remnant]

i have so much to explore on just one discord server .. overwhelmingg

[keen hollow]

whdya mean?

[nova remnant]

countless other channels under one server

i haven't yet explored all the channels in this one

[keen hollow]

oh

well go check em out

[nova remnant]

i will have to make a schedule to do that

trying out the bpn model

bright proto nuke that eface merged

beautiful girl, short summer outfit, necklace, autumn leaves, blue sky, puffy clouds.

SDXL

its got better effect than sd1.5

and on top of that i think eface did some tweaking too

[toxic dragon]

https://cdn.discordapp.com/attachments/453649281744764938/1181080722941956196/00109-3039924041.png?ex=657fc223&is=656d4d23&hm=f151c6745a76ac51f7a301152aedc45e4ebea0ec21195b014377a45779678399&

[nova remnant]

BPN

[nova remnant]

nice... however sdxl models need a lot of training still

you could say that, pretty much all text to image apps started out around the same year .. late 2021 -or early 2022

sure

break shoudl work but im not sure if that works with placing objects

i use that mostly for colors

what you are trying to do is regional prompting... there are extensions for that

idk about comfy .. lol but in a1111 there is

[native halo]

yes,gettin that on sd is very hard

its easy if u inpaint it

or img 2 img

idk what happened here but i support the annual banana/hatsune miku festival

our new god

[nova remnant]

i can't atm, i dont have the extension installed

but let me get you something instead

https://youtu.be/WwATqUp5gBc

big changes ahead #🍥｜anime message

it already does

what issues are you having with images on a1111

example?

ok

comfyui and a1111 are not the engine btw, they are interfaces for the same SD algorithm, just that they have different method of doing things

they are both gui

[vital raptor]

that is the gui

[nova remnant]

ahh well they have different approach but they both uses the same sd engine

they may differ feature wise to some extent

just another inteface

[vital raptor]

it's just a different interface for stable diffusion

[finite herald]

New?

[nova remnant]

you didn't clone the git?

im not good with troubleshooting it, but if you are stuck you could ask in #🤝｜tech-support or you could use this guide to fresh install a1111....

https://www.greataiprompts.com/guide/install-stable-diffusion-in-windows/

a gothic girl on a bed of roses

[nova remnant]

a gothic girl in a tub of milk take caution.......

[vital raptor]

lol, if i leave the 'take caution' in the prompt it puts her clothes back on

a gothic cat girl sitting in a dark forest

generating 2048x2048 images with TensorRT + LCM is fast, but it takes a bit over 21GB of vram. If I'm doing anything else at the same time it gets really slow 😄

[nova remnant]

21gb .... crazy load

[vital raptor]

yeah, 21.6GB, but 7.4-8.3s

[nova remnant]

so you are able to use LoRA with Tensor RT?

was asking Varghoss 🙂

[vital raptor]

kinda, normally you have to convert it to the TRT profile and it has limitations, like only 1 lora at a time

and you have to generate the engine for every model + lora combo

[nova remnant]

yeah

[vital raptor]

but, the lcm lora doesn't count as a normal lora, so it doesn't need to be converted

[nova remnant]

i see

although i only use LCM LoRA mostly on sdxl

[vital raptor]

a pastel gothic cat girl sitting in a dark forest 🙂

I ended up merging my custom lora with my bluepencil merge so I don't have to use a different RT engine for it, not having the extra lora loaded saved me about 0.5 seconds per gen at it's fastest

[nova remnant]

nice idea

[vital raptor]

using 1.5x instead of 2x on the hires fix drops it down to 4.9-5.3s and only uses 16GB vram

I've noticed using LCM isn't very random though, not a lot of variation in the renders unless you change the prompt

[nova remnant]

lcm adds some level of coherence i think

[nova remnant]

Meow

[nova remnant]

Might be the model?

do you have --xformers added?

if you are using nvidia that is

[vital raptor]

yeah

[nova remnant]

edit the file in notepad webui-user.bat

got to restart the sd now

3060

[vital raptor]

Could be, bluePencil seems to be pretty consistent. Could also be the difference between 1.5 and sdxl, haven't used sdxl much until now

[nova remnant]

depends, but in average 1.5it/s

@wispy canopy something i should mention ... bikini top and jogger's pants makes the character show up only in bikini and panty

[wispy canopy]

xD 😦

thats'no good

[nova remnant]

minor issue 🙂

but could be nice to fix that

none other than that

[wispy canopy]

?

[nova remnant]

yes would liek that

other models can do that

[wispy canopy]

that is my model

just now

[vital raptor]

'jogging pants' might work better than 'jogger pants' also

[nova remnant]

what was your prompt?

[wispy canopy]

are you prompting "jogger pants" instead of "jogging pants"?

[nova remnant]

yes

[vital raptor]

on my model at least it makes a difference

[wispy canopy]

because one of those two isn't a thing

call them jogging pants -> it'll work

[nova remnant]

let me rephrase that

doenst

a girl with long brown hair, wearing bikini top, jogging pants, in a lush paradise, palm trees, crystal-clear waters, ancient structures

[wispy canopy]

lets see

[vital raptor]

a cat girl in bikini top and jogging pants

[wispy canopy]

not sun -> nuke

[nova remnant]

beaut

[wispy canopy]

@nova remnant, it's the crystal clear waters that's pushing the latents towards a full bikini

[nova remnant]

lol oh snap

[wispy canopy]

@nova remnant : a girl with long brown hair, wearing (bikini:0.2) top, jogging pants, in a lush paradise, palm trees, crystal-clear waters, ancient structures

[nova remnant]

ahh so the weight fixed that .. now i got to deal with jogging pant to show up right instead of shorts

[vital raptor]

if there's a significant speed difference it's either not doing the same thing or it's not setup right

[wispy canopy]

?

[nova remnant]

yeah geting that

want her in full pant

[nova remnant]

this

[wispy canopy]

then prompt it, silly!
anime flat color absurdism , a girl with long brown hair, wearing (bikini:0.2) top, full jogging pants, in a lush paradise, palm trees, crystal-clear waters, ancient structures

FULL jogging pants

[vital raptor]

with a lowend card you might need something like --lowvram or --medvram, there might be others but i don't know

[steady grail]

comfy is good only when finely tuned for your hardware

[wispy canopy]

comfy is good in nearly all cases, just a bit less intuinitive

once you get used to it, it's not bad at all

[steady grail]

with the wrong settings, the results are very jerky

[nova remnant]

something else might be interfering... i used that exact prompt this time

[wispy canopy]

both 🙂

[nova remnant]

are you using lcm lora btw?

[wispy canopy]

yeah, lcm lora + sampler

you can't beat the consistency it makes

[nova remnant]

odd cause she's showing up in shorts still

[vital raptor]

comfy was too much bother to mess with workflows all the time, i'm rather on improving the actual work, but it has it's uses

[steady grail]

yesterday i got good set up with lcm that i need to test asap

[wispy canopy]

@nova remnant 3/3 she's having the full pants

high cfg

[nova remnant]

ahh i see now, its the TI that was interfering

[wispy canopy]

BREAK... doesn't work in comfy, does it?

[nova remnant]

good find

im gonna stop using that ti

[wispy canopy]

ti?

[nova remnant]

textual inversion

[wispy canopy]

making mulitple condition groups and combining those

long story short -> no

comfy simply interprets prompts differently

there is some nodes that try to replicate A1111 prompting behaviour

[vital raptor]

I think i saw somewhere that you could use multiple text encode nodes to get kind of the same affect, but it's not quite the same

[steady grail]

yep

[wispy canopy]

don't think it's the same tho

[vital raptor]

not sure, i saw it mentioned but I haven't used it at all, and haven't used comfy in awhile

[wispy canopy]

i think you'd need to combine to get an average, like how a1111 does it

[steady grail]

setting concat well is a little painful in my already packed workflow lol

[wispy canopy]

here i am, simply just prompting xD

[steady grail]

things are becaming very big

[steady grail]

stop to prompting flukes!

i know, i know, i'm an expert

i'm deadly sick of stories with Mc without common sense

i remember anime like Sorcerous Stabber Orphen where the protagonist was cool because he know how to deal with his life, not because he was stupidly op, but totally ignorant

anime of the past

the true edgelord of the heroes

i agree

the first fma was longer and more expositive, but brotherhood was a good condensation andrielaboration of the story

the first time i watched it, i was totally mind blown

fr?

yes

[vital raptor]

had to try out BPN

[steady grail]

fr is for real?

or an anime?

[vital raptor]

model

[steady grail]

i'm a litte confused fma:b is full metal alchemist brotherhood, but what means fr?

ohhh ok

now i'm in line

i remember i watched fma when i was still at school, i got influenced by this sense of justice, only to learn later that things are not always equivalent

even in the anime, the weight of a living heart is massive

and things changes, under his influence

the weight of the feelings

that's what value so much the human life

it's not explained too well in the anime even if it tryed hard, but human transmutation was a taboo, because the weight of human feelings was too heavy and no material thing could overweight it

something i learned with experience

and art helped me a lot to understand this

uahahah

[lyric zealot]

clearly nsfw

[steady grail]

you will get banned

[lyric zealot]

it's dangerously cute, too cute in fact

[steady grail]

better not ping sunny or he will get punished for sure

[lyric zealot]

[vital raptor]

with and without my lora

[bitter latch]

#1 merge (with my soft anime style) I call it fuzzyanime

[steady grail]

wow, details re more sharp

what is this lora?

[bitter latch]

2nd merge same style

hmm?

[steady grail]

prompted the monna lisa?

[bitter latch]

prompts mona lisa

[steady grail]

damn

[bitter latch]

I think it dont' work

[steady grail]

too much cfg?

[bitter latch]

tries euler

hmm so far not good merge

[steady grail]

oh, the merged is baked

[bitter latch]

yeah....

so rpg plus my fuzzy anime and protogenanime.. not good idea

or maybe it's the baked in vae

tries again

[steady grail]

lcm always mess up my prompt

[vital raptor]

i don't think it should

it doesnt change mine at all

[bitter latch]

ok merging Fuzzyanime, realcartoon3D and the protogen anime.. this should be.. umm interesting

[steady grail]

without lcm, i prompt a fantasy landscape and i get a landscaper with a big tree, i use lcm and i get instead of a tree, a big face

ok, the result is beautiful, but it's not what i prompted

it's an evident bias from the model

do you want a suggestion?

[bitter latch]

goes out there and tries LCM with the new merge

[steady grail]

stop promptng low quality 🤣

[bitter latch]

ooooh

[steady grail]

better suggestion, upscale the results and prompt quality prompts like hdr, masterpiece and more

[bitter latch]

hmm

centered, Miku Lisa Painting, art by Leonardo da Vinci and Ghibli, sharp focus, elegant detail, HDR, detailed shadows, cinamatic lighting, unreal render, realistic, fine art illustration, 3.5D, elegant shading, <lora:Ink scenery:0.5>magical, Vivid multi-colored Ink's, detailed outline's, lightly shaded

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1181138636561989652/00128-4049474801.png?ex=657ff813&is=656d8313&hm=e628e8d63db47c7a15774f1c9d079e49ce167cf59b691dee2ad97c0247f27aef&=&format=webp&quality=lossless&width=441&height=662

[bitter latch]

centered, Miku Lisa Painting, art by Leonardo da Vinci and Ghibli

doh

what?

[steady grail]

why i get this big face, instead of a tree?

but i must admit, it's very cool

but it's not what i prompted!

[bitter latch]

what did you prompt?

[steady grail]

||(masterpiece, best quality, highly detailed:1.2), made by miyazaki
2d game scene

HDR photo of landscape by Gustave Moreau,Thomas Kinkade,James Gurney. Carne Griffiths. Frank Frazetta. van gogh,Alberto Sevesooil paint,masterpiece,Realistic,deep colors,blue tint,only bronze gold moon,night scenery,fields,Field,Intricate,detailed,sharp,clear,Better image quality,harsh brush strokes,High dynamic range, vivid, rich details, clear shadows and highlights, realistic, intense, enhanced contrast, highly detailed||

[bitter latch]

Tries with breaks